← sources

Release Paperless-AI v4.1.0 and a few words to the community

source: reddit (reddit:selfhosted)published 2026-11-17 11:59ingested 2026-11-18 02:39

retained text

**Hey fellow selfhosted-gang,**

Today I want to address a few words to the community regarding Paperless-AI and its repo health. As many of you know, there was a seriously long absence from me and the project itself. I want to reach out and explain what kept me from maintaining this project (in a TL;DR):

Mental health issues -> marriage -> first child born -> STRESS -> MORE STRESSED -> "Wait! Babys learn to walk! SUPER STRESSED! :D"

No, for real - time to program was rare, and I valued the time on the couch over maintaining this project. But in early January 2026, I got on the community Discord and announced that I would rewrite the entire codebase, as it had become a critical, unmaintainable spaghetti mess over time. At first, I was super confident I could deliver the final product quite early, but as time flew by, I realized this was much more work than I could handle in a few evenings.

So, long story short - I used AI more frequently this time to help release a working and modern solution. But even with the help of AI, it took months of manual coding, prompting and debugging.

By the time I was done and ready to release, a MAJOR security flaw was found in the old versions (≤ 3.x) that led to a CVE score of 9.8. Just days before I wanted to release the new version. Thanks again to the reporter of this security issue <3 I never thought that one of my projects would become so widely used (over 7.5 million downloads over the years) that threat actors would exploit this flaw out in the wild. Literally, it never crossed my mind... I learned my lesson.

That was the tipping point to release the new version ASAP. So here we go -> Version 4 is out now.

**Is Paperless-AI even necassry nowadays?**

I would say so, yes! I was usure the whole rewriting process with the upcoming release of paperless-ngx v3 and it's included AI features. But after testing it I was keen that a seperate solution will fit better for most people. Me personally - It really disappointed me big time.

**What is the benefit over paperless-ngx (AI integration), paperless-gpt, zettelrobbe (fork of paperless-ai) and other fellow competitors?**

It depends – they all have their reason to exist. ;)

# Spotlight for Paperless-AI:

**What's new in 4.1**

* **OCR with PaddleOCR (optional):** scans, photos and documents with garbled Tesseract text are read locally with PaddleOCR (PP-OCRv5 via ONNX Runtime). No extra container and no cloud service; the models (\~20 MB) are downloaded on first use.
* **Runs before the AI:** automatically for documents *without text*, with *poor text* (there's a text-quality score) or *always*, or only when you start it. Titles and tags then come from good text.
* **New OCR page:** every document with its text-quality score, filters for missing/garbled text, bulk OCR, and Paperless text vs. OCR result side by side. The AI re-names and re-tags the documents afterwards.
* **Better text in Paperless too:** the OCR result can replace the document text in Paperless, which also improves the Paperless full-text search. You can undo it.
* Straightens pages scanned sideways or upside down. Covers 30+ Latin-script languages plus Cyrillic, Greek, Arabic, Chinese/Japanese, Korean, …
* **CPU or GPU:** about 1–2 s per A4 page on a 4-core CPU. New image `clusterzx/paperless-ai:latest-cuda` for NVIDIA GPUs (CUDA 13, driver ≥ 580, falls back to the CPU). `latest` stays the small CPU image for amd64 and arm64.
* **Prompt wizard:** answer a few questions instead of writing a prompt yourself:
   * your own name, so it's never used as correspondent (the most common mistake)
   * output language
   * title style or a fixed pattern
   * how many and what kind of tags
   * short company names, document types and your own rules
* You get a live preview and can test it on real documents before saving.
* **Privacy:** ONNX Runtime's usage telemetry is switched off (it's on by default upstream).

**What's new in 4.0 (in case you missed it)**

* **Ask your archive (RAG), rebuilt:**
   * chunked passages and hybrid BM25 + vector search
   * streamed answers with cited sources and follow-up questions
   * filters (date, correspondent, type)
   * "how many …?" answered with exact numbers from Paperless
   * incremental index updates
* **Embeddings:** built-in multilingual model (offline, CPU), Ollama, OpenAI, Azure or any OpenAI-compatible API, or keyword-only.
* **AI providers:** OpenAI, Anthropic (Claude), Ollama (fully local), Azure OpenAI and any OpenAI-compatible API (DeepSeek, OpenRouter, LiteLLM, vLLM, LM Studio, Gemini …). Thinking models are supported.
* **Paperless-ngx 2.x and 3.x:** the API version is negotiated automatically.
* **Settings apply instantly** – no more `.env` editing and container restarts.
* **Real undo:** the history restores the original values in Paperless.
* **Processing:**
   * custom fields (text, number, monetary, date, select …)
   * trigger tags
   * restrict the AI to existing tags/correspondents/types
   * external API context
   * rate limits pause the queue instead of failing documents
   * instant processing through a Paperless workflow webhook
* **Tools:** document chat, manual review, prompt playground, dashboard, logs & diagnostics.
* **New UI:** light/dark mode, accent colours, Ctrl/⌘+K command menu, works on mobile.
* **Upgrading from 3.x is automatic:** settings, user, processing state and history are migrated on first start.

Link to GitHub:  
[https://github.com/clusterzx/paperless-ai](https://github.com/clusterzx/paperless-ai)

Cheers, Clusterzx

document fetched 2026-11-18 00:30 · canonical https://reddit.com/r/selfhosted/comments/1xyim5h/release_paperlessai_v410_and_a_few_words_to_the